In this talk, we present two examples of how network-based approaches can be used to improve robustness and insight in both artificial intelligence and biological data analysis. First, we explore how networks of interacting Large Language Models (LLMs) can be used to mitigate a critical vulnerability: position bias in multiple-choice question answering. This position bias is consistent across model types and sizes, with accuracy swings as large as 55% observed in models like Falcon-7B. To address this, we introduce an Opinion Dynamics algorithm, which models a decentralized network of LLMs that update their beliefs based on local peer interactions. Second, we turn to the analysis of high-dimensional omics data, where sparsity and noise challenge conventional machine learning. We propose a graph-theoretic framework to infer statistically enriched co-occurrence networks from small-sample 16S microbiome data. These networks preserve biological meaning while enabling dimensionality reduction and improved interpretability. We further demonstrate how optimal transport can embed omics data into latent spaces that enhance classification and patient stratification. Together, these examples illustrate how network structures—whether composed of intelligent agents or biological entities—can be leveraged to improve prediction, resilience, and interpretability across domains.

Local Event Organiser: Georgios Karagiannis
Local Group Secretary: Louis Aslett